In most cases, a suitable solution to a problem cannot be achieved by using a single neural network. This is indeed not the way to achieve the best performance and it is generally good practice to break down the overall problem into smaller steps. Deepomatic workflows give you this capacity.
You can create complex solutions without having to worry about deployment or runtime.
A workflow corresponds to a directed graph without cycles. It is defined with a YAML file where you need to list all the data processing steps required in your solution.
The workflow.yaml file defines:
The entries and outcomes of the workflow
The structure of the steps
The configuration of each step
The order in which you write the steps doesn't really matter. The workflow server takes care of reconstructing a graph from the inputs of each step.
Every workflow file starts with mandatory metadata: the workflow configuration version and the name of the workflow.
Workflow structureversion: "1.2"workflow_name: My first workflowworkflow:entries:### List of your entriesoutcomes:### List of your outcomessteps:### List of your steps
The entries defined can be later used as an input for other steps. An entry is composed of two mandatory fields: a name and a data type.
Workflow entriesentries:- name: image inputdata_type: Image- name: contextdata_type: Text
The data type must be one of the following three: Image, Text or Number.
Outcomes are optional and are especially useful when you want to build an augmented technician API. They are composed of the following fields: a name, an output_of step name, a concept, and optionally a regions field.
Workflow outcomesoutcomes:- name: hello worldoutput_of: hello world stepconcept:name: speech to the worldtype: Text
The outcomes must correspond to the list of checkpoints that you want to enforce for the technician on the field, and this is also the information that you will be able to display on the technician application.
In addition to this information, it is possible in a native way to add the visualisation of any objects that may have been detected (bounding box), to provide more information to the technician in the field.
Workflow outcomes with regionsoutcomes:- name: hello worldoutput_of: hello world stepconcept:name: speech to the worldtype: Textregions:- small_object_detector- big_object_detector
In the above example, all bounding boxes from the
big_object_detector steps are passed in the outcome (see below for the inference steps). The good practice is to list here all inference steps corresponding to an object detection task that are useful to explain the final prediction of the checkpoint.
Steps allow you to build your analysis graph. There are several steps available by default, but you can also write custom steps in python if the operation you need to perform is not implemented.
The business logic is detailed via those steps which are listed one after the other (the order does not matter). A step is composed of the following fields: a name, a type, some inputs, and args. The args depend on the type of the step.
Workflow stepssteps:- name: my first steptype: Inferenceinputs:- image_inputargs:model_id: 12345concepts:- persons
Below are listed all steps that are available by default in the Deepomatic library.
An inference step lets you use any model trained on the Deepomatic platform. When writing a workflow, you only refer to a model via its id, and not to a model version (see here for the difference between a model and a model version).
Inference stepsteps:- name: my first inference steptype: Inferenceinputs:- image_inputargs:model_id: 12345concepts:- my objects
inputs should be an image as this is the only input supported for now by the Deepomatic models.
concepts is a name given to the output of the step and that you need to use this output in other steps.
The PredictionRouter step divides the list of input regions according to the name of your model's concepts in the previous inference step. This step therefore necessarily follows an inference step.
PredictionRouter stepsteps:- name: items_routertype: PredictionsRouterinputs:- detectorargs:routing_name: itemsrouting_values:- PLATS_PRINCIPAUX- BOUTEILLES- FRUITS- PAINtop_prediction: False
routing_name corresponds to the concept name of the inference step preceding the current step. The
routing_values must be coherent with the concept names that your model has been trained on. A PredictionRouter step is then only compatible with a specific set of models.
The last argument
top_predictionis optional and lets you filter even more the regions generated by the preceding inference step, by using only the top prediction when relevant. By default,
To use the output of a PredictionRouter step as input for another step, here is the syntax that you must use:
PredictionRouter step - Use output as inputsteps:- name: detectortype: Inferenceinputs:- image_inputargs:models_id: 23456concepts:- items- name: items_routertype: PredictionsRouterinputs:- detectorargs:routing_name: itemsrouting_values:- PLATS_PRINCIPAUX- BOUTEILLES- FRUITS- PAINtop_prediction: False- name: fruits_classificationtype: Inferenceinputs:- items_router:2args:model_id: 54321concepts:- fruits type
In the above example, regions are created with the
detector step. The
items_router step then creates branches for the different categories of items detected. Finally, for one of those categories,
FRUITS, the bounding boxes are sent to another model to determine the type of fruit.
The RegionCounter step gives you the capacity to count regions and it therefore only accepts a list of regions as input.
RegionCounter stepsteps:- name: furniture_countertype: RegionCounterinputs:- furniture_detectorargs:entry: furniture_imageconcept_name: furniture_countcount_only_concepts: [furniture]
The count number if stored in a new region attached to the specified entry.
It is possible to add the optional argument
count_only_concepts to filter the regions.
Finally, it is possible to write custom steps in Python to implement the steps that are missing to build you specific workflow. To do so, you need to write the code of those custom steps in a separate Python file
For each custom step, you need to create a new class that inherits from the CutomNode class, and implement the
process methods. You will then be able to add steps with this type in your
workflow.yaml file (the name of your class is the type of the step).
Custom stepfrom deepomatic.workflows.nodes import CustomNodeclass MyCustomStep(CustomNode):def __init__(self, *args, **kwargs):super().__init__(*args, **kwargs)def process(self, context, *custom_step_inputs):### Implements the logic of your custom step
Let's look at the following custom step and see how you can add arguments and inputs in the logic you implement.
steps:- name:type: MyCustomStepinputs:- image_input- categoryargs:my_first_arg: context_technicianmy_second_arg: intervention_type
You might indeed need to add arguments to your custom step, so that you can use the same custom step and adapt it to a specific usage (the
model_id is for instance an argument of the Inference step). To do so, you need to specify those arguments in the signature of the
__init__ method. You will also probably want to save those values as attributes to use them in the
def __init__(self, *args, my_first_arg, my_second_arg, **kwargs):super().__init__(*args, **kwargs)self._context_technician = my_first_argself._category = my_second_arg
In the same way, your step will use input that might be entries of your workflow or outputs from other steps. You are able to use those inputs in the process method to implement the logic that you need.
def process(self, context, image_input, category):
process method should return either:
the list of modified or created regions
an empty list is nothing has been created or modified
None if you want to stop the execution of the following steps in the corresponding workflow branch